Repository for experiments on evaluating the conversations modelled as episodic knowledge graphs, according to the @Leolani framework
In the data
folder you wil find a folder per conversation setup. Inside each folder there are several scenarios. The structure
of each scenario is the following:
Folders | Description |
---|---|
\{DATESTAMP} | RDF .trig files per utterance that went in the brain |
\automatic_evaluations | CSV file containing all proposed graph metrics and the aggregated human annotations and baseline automatic metrics |
\correlations | CSV files containing correlations between automatic metrics, and human annotations |
\human_evaluations | CSV files containing human annotations and baseline automatic metrics |
\plots | Plots for correlations and conversation flow |
In the src
folder you will find the following:
Folders | Description |
---|---|
\dialogue_creation | Code to generate episodic knowledge graphs for human-human dialogues in the MELD dataset |
\dialogue_evaluations | Code to average and correlate human and automatic annotations |
\graph_evaluations | Code to recreate conversations through RDF files, and compute metrics about the graphs |
This repository uses Python >= 3.7. The following is the recommended set up for this project.
conda create --name evaluating-coversations-as-ekg python=3.7
conda activate evaluating-coversations-as-ekg
pip install --upgrade pip
pip install -r requirements.txt --no-cache
python -m ipykernel install --name=evaluating-coversations-as-ekg
To rerun the graph metric calculations, run src/graph_evaluations/evaluate_rdf_scenarios.py
using one of the available configurations
in resources/running_configs.txt
. To only recreate the plots, run plot_rdf_scenarios.ipynb
.